A label is not a limit

Three days ago the European Commission started enforcing the part of the AI Act that touches nearly every support chatbot in Europe. From 2 August, an AI system that talks directly to a person has to make clear it is an AI, unless that is already obvious from context. The ceiling for getting it wrong is 15 million euros or 3 percent of worldwide turnover.

For most support teams, that is the whole of it.

The same package that brought the label forward pushed the substantive rules back. Standalone high-risk systems now have until 2 December 2027, and AI embedded in regulated products until 2 August 2028. And an ordinary support agent, one that answers questions about orders and policies and account settings, mostly never lands in the high-risk category anyway. That list is about creditworthiness, eligibility for public benefits, life and health insurance pricing, recruitment. Answering someone about a late parcel is not on it.

So for the thing most of our readers run, European law now asks exactly one question, and it is a question about identity. What is this? Not: what is it allowed to do to my account.

The two-agent problem

Picture two support agents on two websites this morning. Both open with the same line. You are chatting with an AI assistant.

The first can look things up. Order status, policy text, where a package is. Anything that changes a record goes to a person.

The second can issue refunds up to a number it was told about once, change the shipping address on an open order, apply account credit, cancel a subscription. It picks which of those to do by reasoning about the conversation in front of it.

Same disclosure. Same compliance posture. Same exposure to that 15 million euro ceiling. From where the customer sits these are not remotely the same product, and the label they both display carries none of the information that separates them.

None of which makes the law wrong. Telling people they are talking to software is worth having as a floor, and it was plainly the piece that could be agreed on fastest. The trouble starts if a support team decides the sentence at the top of the chat window is its AI accountability story. That is the cheapest question on the sheet, answered.

The second sentence

The disclosure that would actually mean something is the one about authority. What is this agent allowed to execute on its own, and what does it only get to propose?

That question has an answer you can write down and check. It does not depend on the model behaving. It depends on the decision living somewhere the model cannot reach.

That split is the shape of IMCeleste. Celeste can draft any reply and suggest any action at all. Whether the action runs is settled by deterministic code against limits the owner set, so a refund ceiling is a ceiling and not a strongly worded request. Authority then arrives one topic at a time. Celeste practices a topic against real historical tickets, the owner reads what the answers would have been, and only that topic goes live. Shipping questions can be autonomous in week one while anything touching billing stays in draft for a quarter. That is how it works.

The side effect worth having is that it gives you something real to say when a customer asks. Not “we use AI responsibly,” which means nothing. Something closer to: this agent can tell you where your order is, and a person signs off on anything that touches your money.

Where the bar goes next

December 2027 arrives, harmonised standards land, and the high-risk regime starts asking operators to demonstrate human oversight, logging, and documented risk controls. Any team that spends the gap wiring up a disclosure banner and nothing else will do that work under deadline, on a live agent, with customers attached to it.

A team that already knows which topics its agent is trusted with, and can produce the record of who granted that trust and when, will mostly be transcribing what it already does.

The compliance answer and the good answer are a long way apart right now. That gap closes on a schedule that is already published. Start a trial and see what the earned version looks like.